On this page you can download the discussion paper that was submitted for publication in the Journal for Media Linguistics. The blogstract summarises the submission in a comprehensible manner. You can comment on the discussion paper and the blogstract below this post. Please use your real name for this purpose. For detailed comments on the discussion paper please refer to the line numbering of the PDF.
Blogstract of
No choice. On the stylistics of AI-generated texts
by Simon Meier-Vieracker
In public discourse surrounding artificial intelligence (AI), the style of texts generated by applications like ChatGPT is often critically examined. Observations frequently include that they sound „too perfect,“ „eerily smooth,“ or lack „a distinct voice“ or „emotional depth“. These comments point to linguistic features, the analysis of which falls under the field of stylistics. But how precisely can we grasp the „style“ of AI-generated content, and what do these capabilities reveal about the underlying mechanisms of Large Language Models (LLMs)? In this paper, I seek to explore these questions.
Many existing stylistic studies on AI texts employ a reductionist concept of style, often limited to countable linguistic features for authorship detection. Instead, I propose applying sociolinguistic and pragmatic style theories, which define style as a meaningful choice among alternatives – a „socially relevant (meaningful) way of performing an action“. This choice generates additional meaning, as what is said always exists within a ‚horizon‘ of other possibilities that were deliberately not chosen.
An experiment demonstrates the remarkable stylistic abilities of LLMs. Three models (ChatGPT 4o, Claude Sonnet 4, Google Gemini 2.5 Flash) are prompted to rewrite a short narrative text into 13 different styles. These styles ranged from „conversational,“ „formal,“ and „emotive“ to „academic“ and „tabloid“. The models executed the task effortlessly, producing texts that differed significantly from the original but maintained consistent stylistic features within each given style. A stylometric cluster analysis confirmed that texts of the same style – even from different LLMs – grouped closely together. For example, the „conversational“ texts consistently showed discourse markers typical for oral narratives, tag questions, direct listener addresses, and interjections, all indicating a dialogical and emotionally engaged speech situation.
Despite this impressive flexibility and style imitation, LLMs do not make choices in the human, meaning-generating sense. They are mechanistic, probabilistic systems that predict the next word in a sequence based on statistical patterns learned during training. Their generated „meaning“ is a „dumb meaning,“ which is ‚parasitically‘ dependent on a human interpreter. LLMs lack the „horizon of an ‚and so forth‘ of action and experience“ that humans possess in their decisions.
How, then, can the success of LLMs in style imitation be explained? The key lies in their training. LLMs are trained on vast amounts of human texts, which contain not only grammatical patterns but also complex patterns of human stylistic choices and metapragmatic categorizations. Metapragmatic references describe how humans themselves typify and categorize styles, such as when we speak of „formal language“ or „tabloid style“. These everyday style labels and their associated stereotypical descriptions are statistically learned by LLMs. When an LLM is prompted to write in a specific style, it reactivates these learned patterns, effectively reproducing „intelligible textures“ it has learned from human texts.
In summary, LLMs master the core principle of style – expressing the same thing in different ways. However, they do not make human choices. Instead, it is the humans who, through our prompts, have the choice to have them write in diverse and interpretable styles, by leveraging the human stylistic and metapragmatic patterns represented within the models. This highlights the fundamental role of metapragmatics in human language use and its indirect influence on the capabilities of AI-generated texts.
(This blogstract was AI-generated by NotebookLM and post-edited by the author. The prompt included the full maniscript and the journal’s definition of blogstract as „a specific abstract form in which the relevance of the study and the essential contents of the discussion paper are summarised in a generally understandable way and made comprehensible for both interdisciplinary and non-academic discourse“)